ClownGuide

This is the RES4LYF node people fall in love with, even more than the samplers. ClownGuide lets you feed a reference latent into sampling and steer the result toward it - composition, color, layout - without a ControlNet or a fixed img2img denoise. One longtime user put it well: it lives "in the vicinity of img2img with high-ish denoise, or unsampling / RF inversion / ad-hoc ControlNet for models that don't have controlnets." That's the pitch. If your model has no ControlNet ecosystem (a lot of the newer flow-matching ones don't), this is how you guide it anyway.

It's also the backbone of the pack's style-transfer and faceswap tricks. You point it at an image and tell the sampler "head in this direction," and how hard, and for how long.

How it works

You hand it a guide latent - usually your reference image, VAE-encoded - and it produces a GUIDES object that plugs into a Shark/ClownsharK sampler. During sampling it nudges the trajectory toward the guide instead of pinning it. Because it works inside the sampler rather than as a pre-baked init, you control when the guidance applies, so you can lock composition early and let the model invent detail late.

The inputs and outputs that matter

Lots of knobs, but a beginner only touches a handful:

• guide (LATENT, optional but the whole point) - your reference, encoded to latent.
• guide_mode (default epsilon) - how the guidance is injected. epsilon and flow are the two most people actually use; the author and power users flip between them. The rest (pseudoimplicit, sync, inversion, the self_refine pair) are stronger/experimental variants - leave them until you know why you want one.
• weight (default 0.75) - strength. Higher sticks closer to the guide, lower gives the model more freedom. This is your main dial.
• start_step / end_step (defaults 0 / 15) - the window the guide is active. Guiding only the early steps locks composition then frees detail; that's the trick for "same layout, new content."
• weight_scheduler (default beta57) - how weight ramps across that window. beta57 is a sane default; it's the same beta57 the rest of the pack loves.
• mask (MASK, optional) - restrict guidance to a region; invert_mask flips it.

Output is guides (GUIDES) → into the sampler's guides input. There's no image out here; the effect shows up when the sampler runs.

How to install it

ComfyUI Manager: search RES4LYF, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ClownsharkBatwing/RES4LYF
cd RES4LYF
pip install -r requirements.txt

Portable ComfyUI: use your embedded pip. Restart, then hard-refresh with F5. Installing rgthree-comfy alongside gives you the tidy nested menus the README recommends.

Common issues & troubleshooting

The guide overpowers everything. Drop weight, or shorten the window with a lower end_step. A full-strength guide across all steps is basically a copy machine.

The prompt is secretly doing the work. The README's own advice: run once with the guide off (or weight at 0) to confirm the change you're seeing is actually the guide and not your text prompt. Easy to fool yourself here.

Model doesn't support guidance. Guides need the matching patcher for your architecture - ReFluxPatcher for Flux, ReWanPatcher for Wan. No patch, no effect.

Which guide_mode? Start with epsilon or flow. If composition guidance feels weak or smeary, try the other before touching the experimental modes - the difference between epsilon and flow is usually bigger than anything the fancier modes buy you.